Related Experiment Video
Updated: Aug 5, 2026
10:24
Neutron Radiography and Computed Tomography of Biological Systems at the Oak Ridge National Laboratory's High Flux Isotope Reactor
Published on: May 7, 2021
Artificial neural network Radon inversion for image reconstruction
A F Rodriguez1, W E Blass, J H Missimer
1Department of Computer Sciences, Instituto Tecnologico y de Estuidos Superiores de Monterrey, Mexico City, Mexico.
Medical Physics
|May 8, 2001
Summary
This study explored using a back-propagation neural network (BPN) for positron emission tomography (PET) image reconstruction. The BPN successfully reconstructed images from arbitrary objects after training on Gaussian images.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computerized Tomography
Background:
- Image reconstruction is crucial for computerized tomography (CT).
- Filtered backprojection (FBP) and algebraic techniques are common.
- Positron emission tomography (PET) requires accurate image reconstruction.
Purpose of the Study:
- To investigate the feasibility of applying a back-propagation neural network (BPN) for tomographic image reconstruction.
- Specifically, to assess BPN performance in positron emission tomography (PET).
Main Methods:
- A feed-forward back-propagation supervised artificial neural network (BPN) was designed and trained.
- The network was trained using Gaussian test images.
- Performance was evaluated based on reconstruction accuracy from projection data of arbitrary objects.
Main Results:
- The BPN successfully reconstructed images from projection sets of arbitrary objects when trained with Gaussian images.
- Optimal network design includes middle layer nodes significantly fewer than input/output nodes.
- Training iterations decreased exponentially with increased middle layer nodes.
- Optimal reconstruction accuracy was achieved with a Gaussian FWHM of three pixels for single-width training sets.
Conclusions:
- The BPN demonstrated feasibility for general image reconstruction, independent of training data specifics.
- Current accuracy is insufficient for immediate PET application.
- Future refinements could lead to a network capable of fast 3D image reconstruction from noisy PET data.
